---
pretty_name: SWE-PolyBench Retrieval
task_categories:
- text-retrieval
tags:
- mteb
- rteb
- code-retrieval
- agentic-coding
configs:
- config_name: corpus
  data_files:
  - split: test
    path: corpus/test-*.parquet
- config_name: queries
  data_files:
  - split: test
    path: queries/test-*.parquet
- config_name: default
  data_files:
  - split: test
    path: qrels/test-*.parquet
- config_name: top_ranked
  data_files:
  - split: test
    path: top_ranked/test-*.parquet
---

# SWE-PolyBench Retrieval

Retrieval-formatted version of [SWE-PolyBench](https://huggingface.co/datasets/AmazonScience/SWE-PolyBench), prepared for the RTEB code
retrieval benchmark in [MTEB](https://github.com/embeddings-benchmark/mteb).

Each query is a coding task (a GitHub issue). Its candidate documents are the files of
the corresponding repository snapshot; the relevant documents are the files that the
reference solution modifies.

| | Count |
|---|---:|
| Queries | 2,109 |
| Documents | 12,037,557 |
| Qrels | 5,423 |

## Format

Standard MTEB retrieval layout, loadable with MTEB's default `AbsTaskRetrieval` loader:

- `corpus`: `id`, `text`. The id is `<example_id>/<repository-relative file path>`,
  the text is the file content.
- `queries`: `id`, `text`.
- `default` (qrels): `query-id`, `corpus-id`, `score`.
- `top_ranked`: `query-id`, `corpus-ids`. Every example has its own repository
  snapshot, so each query is evaluated only against the documents of its own example
  (the full list of that example's files, not a pre-ranked subset).

All configs have a single `test` split.

## Source and license

Derived from [SWE-PolyBench](https://huggingface.co/datasets/AmazonScience/SWE-PolyBench) via `MongoDB/rteb-public-agentic-coding`.
The upstream dataset and the included repositories retain their original licenses
and attribution.